Skip to content

ADR-0388: Ingest BVI-CC as the second tiny-AI training corpus

  • Status: Draft
  • Date: 2026-05-02
  • Deciders: TBD
  • Tags: ai, fr-regressor, corpus, license, bristol

Context

fr_regressor_v1 (T6-1a, commit e421d70) trains on the Netflix Public corpus only — 9 reference + 70 distorted clips, single MOS convention, no explicit codec axis. The codec-aware fr_regressor_v2 plan in ADR-0235 needs subjective labels paired with a real codec sweep. The companion research digest Research-0046 walks the Bristol VI-Lab BVI- family and concludes that BVI-CC* is the smallest useful candidate: 9 references × 34 distorted variants across HM, AV1, VTM at four resolutions, with DMOS labels.

The decision is whether to make BVI-CC the first Bristol corpus ingest for the fork, ahead of BVI-DVC (no labels), BVI-AOM (no labels, mixed-licence subset), or BVI-HD (single codec only).

Decision

We will ingest BVI-CC as the second tiny-AI training corpus, behind Netflix Public, in a single dedicated PR. The PR ships: the JSON manifest at ai/src/vmaf_train/data/manifests/bvi-cc.json, a mos_convention column in the feature parquet schema, loader support for inverted-DMOS-to-MOS normalisation, a per-clip licence flag column (carried for forward-compatibility with BVI-AOM), and the user-facing doc page at docs/ai/training-data.md. Raw clips remain gitignored under .workingdir2/bristol/bvi-cc/. The parquet itself is a local-build artifact, not committed.

Alternatives considered

Option Pros Cons Why not chosen
BVI-CC first (chosen) DMOS labels; explicit HM / AV1 / VTM codec axis; small enough (~250–400 GB) to fit on one disk; CC-BY on derived data Form-gated download (~2 day SLA); DMOS↔MOS conversion adds one normalisation step —
BVI-DVC first Largest sequence count (800); widest resolution range; useful for parity soak No subjective labels — does not move fr_regressor_v2 forward; 700 GB – 1.2 TB footprint Wrong corpus for the labelled-MOS goal; can come later for a separate parity-soak PR
BVI-AOM first Authoritative size figure (124 GB packed); clean direct-S3 download (no form) No subjective labels; mixed licence — CableLabs subset is CC-BY-NC-ND 3.0 (no derivatives), needs per-clip licence map Same labelling gap as BVI-DVC plus a licence-complexity tax we don't need on the first ingest
BVI-HD first DMOS labels; simpler than BVI-CC Single codec (HEVC + HEVC-SYNTH); no AV1 / VVC presence Doesn't exercise the codec one-hot the v2 regressor was designed for
Stay on Netflix Public, defer BVI Zero new licence work; zero new disk Blocks ADR-0235 indefinitely; cross-backend soak still has only 3 pairs Defers the actual problem
Pull two corpora at once (BVI-CC + AOM) One round-trip TB-scale storage commitment; AOM licence map needs separate review; harder to attribute lift in v2 Violates the "one decision per PR" principle

Consequences

  • Positive: fr_regressor_v2 gains a labelled codec sweep with three encoders; the parquet schema gains a mos_convention column, unblocking any future MOS-bearing corpus from any lab; docs/ai/training-data.md becomes the single index of what the fork has ingested.
  • Positive: A 10-clip BVI-CC subset becomes available for a weekly cross-backend parity soak (idle-GPU job), expanding the parity surface beyond the 3 Netflix golden pairs without touching the golden-data gate.
  • Negative: Manifest and parquet schema churn — every downstream consumer of the parquet must accept the new mos_convention column. Acceptable cost; the column is additive.
  • Negative: Disk pressure. The dev box must keep ~400 GB free for the BVI-CC corpus root in addition to the existing ~37 GB Netflix root.
  • Neutral: Reproducibility for future contributors requires re-registering with Bristol; the in-repo manifest plus expected SHA-256 list is the audit trail.
  • Follow-ups:
  • DMOS-to-MOS conversion test (places=4) added to python/test/ against a 5-clip frozen subset.
  • Memory entry mirroring project_netflix_training_corpus_local.md once the corpus is local.
  • Decision deferred: BVI-DVC and BVI-AOM ingest. Revisit after measuring v2 lift from BVI-CC alone.

References